A Study Of Object Tracking In Complex Scenes Based On Compute Vision

With the rapid development of computer technology and the large scale application of video capture devices, the research on computer vision has been attracting more and more attention. Object tracking as a core technology in computer vision has been widely used in video surveillance, Human-Computer interaction, medical diagnosis, precision guidance and many other fields of civil and military. During the research and development of object tracking, many excellent algorithms have been proposed, and achieved good performance in some application scenes. However, there are still many problems to be solved for applying these algorithms to track object in complex scenes of real world.According to the need of theoretical research and practical application, a thorough analysis and discussion are made on those existing object tracking algorithms. For problems of object tracking in complex scenes, the main work and results in this dissertation are as follows:1. To solve the problems of occlusion and deformation during object tracking, an improved object tracking method based on local feature model is proposed. The method extracts the local features as basic elements, and builds the object model by weighting the stability of local features with General Hough Transform. With the stable local features as foreground patches, the global color probability distribution are calculated. Then it adjusts the local feature model in turn and optimize the tracking results in order to reduce the tracking error due to occlusion and deformation. Finally, local features and global features constraint and complement with each other, and update online adaptively. The proposed method improves the accuracy and stability of the local feature model, thus ensuring the robustness for occlusion and deformation.2. Multiple object tracking in video scenes are discussed, and the mainstream multiple object tracking methods are analyzed which are based on object detection. The current object detection algorithms’performance are still not ideal in complex scenes and are prone to false detection and miss detection, thus causing the tracking failures. For this problem, an improved multiple object tracking method based on unstable general detector is proposed. The method uses the general object detector based on histogram of oriented gradients, and correct the detection results with background subtraction. Then optical flow method and particle filter method are adopted to track between frames with the context information, and the results are used to optimize data association, thus achieving stable tracking results.3. Propose a pattern analysis method based on density cluster for high density crowd motion. The method firstly extracts feature points of each frame, and obtains optical flow information by tracking with optical flow method. Then make statistics on optical flow information in a period of time to produce optical flow field including motion information. Eliminates the noise and invalid optical flow information by filtering. Finally, make cluster analysis on optical flow filed with cluster method based on density to obtain motion pattern of the crowd. The crowd motion pattern visually reflects the motion state of the crowd target. Furthermore, with the motion pattern as the priori information, an improved particle filter algorithm is given to verify the effectiveness of the motion pattern.4. To solve object tracking problems in unconstrained environment, the tracking methods based on object detection is studied and two fusion tracking frameworks are proposed at different levels. 1)The first level is algorithm fusion. Propose a hierarchical tracking algorithm that constrains detection results with multiple algorithms. The template matching method is used to enhance the stability of tracking, and median optical flow method is used to enhance the adaptation of tracking. Under the constraint of these two methods, the online adaptive boosting algorithm reduces the occurrence of the drift, thus improving the tracking performance.2) The second level is information fusion. Propose a tracking method with multiple information fusion based on Hough random Ferns. Firstly, use local image patches as low level features to train the local feature detector with Randon Ferns algorithm, and vote to the object location with the detection results. Then superpixels are generated by image segmentation as the middle level features and used to build color probability distribution model. With the constraint and optimizing of the color model, the robustness for appearance variation of the local feature detector is improved. Finally, the proposed method combines these two models based on different features in a co-training framework, and fuse the information of object detection, voting estimate and color probability distribution together, thus to achieve better tracking results.The relevant work is supported by National Science Foundation of China (No. 61173091).